It's 7:30 p.m., and you've just hung up on your third sales call of the day.
The client spent fifteen minutes walking you through their pain points. You scribbled a few keywords in a notebook and told yourself: I'll get to the quote tomorrow, once the rest of today's stuff is out of the way.
But "tomorrow" rarely arrives on schedule. Your inbox is full again in the morning, and the quote you promised gets pushed back one more day. By the time you actually sit down to write it, you've forgotten the specific pain points the client emphasized on the call. The quote comes out slow and generic. A week later they finally reply โ and by then, they might already be talking to someone else.
If you run your own business, take on freelance clients, or you're trying to figure out how to actually put AI to work inside a company's daily operations โ not just chatting with it or generating images โ this scene probably isn't unfamiliar. It's also the first case study that the YouTube channel "ๆๅๅฉ | AI" (Harry Lee | AI) walks through, start to finish, in a two-and-a-half-hour teaching video: using OpenAI's Codex to turn a sales call that goes nowhere into a system that generates the quote, gets it signed, simulates payment, and eventually runs on a cloud schedule, on its own.
At the start of the video, Harry Lee introduces himself as the founder of a company that helps small and medium businesses adopt AI automation, and says he runs a community of over ten thousand people. Those numbers are his own claims โ this site hasn't independently verified them, so treat them as background rather than something to take at face value. What's actually worth breaking down is the workflow he builds on screen, step by step, that you can follow along and replicate yourself.
Before You Open Codex, Know These Three Switches โ Or You'll Be the One Burning the Tokens
Codex is OpenAI's coding and automation assistant. You can install it as a standalone client, or run it inside a browser or terminal. The interface itself isn't complicated: a chat list on the left, and a browser or terminal panel on the right that lets you watch in real time which page it's operating on and which command it just ran. What actually determines whether the whole setup is worth the cost comes down to three settings that are easy to overlook.
The first is the permission mode. Codex lets you choose between requiring your approval for every action, and letting it run fully on its own. Harry Lee's habit is to pick the more conservative middle ground โ important actions get flagged for his review first, so the AI doesn't accidentally delete a file it shouldn't have. If you want more speed, you can hand over full control, but the risk is yours to carry.
The second is the model's effort and speed settings. For the same task, you can ask it to think more carefully or run faster โ but both of those choices show up directly in how many tokens you burn. The more thoroughly it thinks and the faster it runs, the more it costs. Harry Lee mentions that his day-to-day work mostly sits at medium-to-high effort, rarely maxed out, for a simple reason: there's no need to burn the most expensive setting on every single task. That's worth remembering, because it's really the old question of "how do you make AI tokens worth it" showing up again in a new setting.
The third easy-to-miss piece is schedule and plugin functionality โ you can have Codex run a task at a fixed time, or connect it to third-party tools you already use so it can operate them directly. These two features are exactly what turns the whole system from something you click through by hand into something that runs on its own.
From One Prompt to a Reusable Skill
Once you've got the interface down, the real substance is how Harry Lee turned that opening problem โ the quote that keeps getting pushed to "later" โ into something reusable.
He first has Codex build three quote-page templates for three completely different industries โ medspas, dental clinics, and air conditioning repair โ each in a deliberately different style. The reasoning is practical: if you only generate one version, you have no way of judging whether it's actually what you want. Generate three clearly different versions, and now you have something to compare, and you can immediately see which elements to keep and which to swap out.
Once a style is picked, the next step is where the real value shows up: he feeds Codex a transcript of a simulated sales call and asks it to generate a complete quote following the chosen template โ a summary of the client's pain points, the proposed solution, an estimated ROI and payback period, and finally an e-signature field with a mock payment screen. Throughout the process, he emphasizes that he never wrote a single line of code by hand. What he did was keep describing what he wanted in plain language, and ask for changes whenever he wasn't satisfied โ until it was right.
Once the template is stable, he promotes it into a named skill with a fixed workflow โ something that, from then on, outputs a finished quote page any time you feed it a new call transcript. He even runs two subtasks in parallel, having Codex handle quotes for two different industries at once, to confirm the skill reproduces reliably instead of it just being a lucky first run.
Getting Automation Off Your Own Computer: Local Scripts vs. Cloud Scheduling
Once the skill is built, there's still a practical problem: if your computer isn't turned on, this "auto-write-my-quote" workflow simply doesn't run.
Harry Lee first demonstrates a local version โ connecting it to a meeting-transcription tool he already uses, so that after every sales call, the system automatically pulls the transcript, applies the skill, generates the quote page, and hands it back to him for one final review before it goes out to the client. This step already cuts out most of the manual data-shuffling, but the moment your computer shuts down or loses connection, the whole chain stops.
To actually get to "standing by 24/7," you have to add one more layer: move the same logic onto a cloud service, set it to run on a fixed cycle โ say, checking every few hours โ to pull new call records, generate the quote, and wait for human sign-off before sending. He admits he didn't understand cloud setup from day one either. His method was to just ask Codex to explain what each setting does, keep asking follow-up questions when he didn't understand, and only proceed once it made sense. That's also a point the video keeps coming back to: you don't need to already know how to code before you're allowed to move a workflow to the cloud.
Two Guardrails Hiding Behind All That Efficiency
By this point, the whole system sounds tempting: a call ends, a quote generates itself, and you could keep extending the same idea into customer service or follow-up collections. But the video includes at least two warnings worth writing down and sticking to the edge of your screen.
First: never paste API keys, passwords, or any sensitive credentials directly into a chat window with an AI. The right approach is to have the AI point you to a file path, then manually paste the key into that file yourself โ a chat log isn't a vault, and you don't want your key sitting inside some conversation history.
Second: even with the whole process automated, he never removed the human-review step before anything goes to a client. The system is designed to hand the generated quote back to him for confirmation first โ only after he approves it does it actually go out. That step might look redundant, but it's the one brake still held by a human hand in an otherwise "unattended" system. What automation saves you is repetitive labor time, not the responsibility of judgment.
One Piece of Content, Many Channels: Repurposing Also Gets Automated
The video shifts scenes in its later half to content marketing: taking a long-form article or newsletter you've already written and breaking it down into short posts and graphics tailored to different social platforms โ one piece of content, many outputs, without the usual effort of starting from a blank page for each channel. In his demo, he pastes the full text of a real, publicly published entrepreneur's newsletter into Codex and asks it to rewrite and generate copy and visuals for three different platforms. The method here is worth learning from, but if you're planning to apply it to someone else's published, copyrighted work, the extent of the rewrite and proper attribution are things that need separate scrutiny โ that's a different risk tier from applying the same technique to your own original content.
A Free Course Billed as "the Most Comprehensive on the Internet" โ What Standard Should You Judge It By?
The whole video is billed as "the most comprehensive on the internet," "an epic masterclass," and it closes with a pitch for his own paid community, including a promise along the lines of "guaranteed to land your first client within 90 days, or your money back." Language like that isn't unusual for teaching-style YouTube content, but it's also a reminder worth sitting with: a free video packed with genuinely useful, actionable steps, and the thing it's ultimately trying to sell you, are two things you can judge separately. The steps themselves you can test and verify on your own. The community size, the money-back guarantee, and similar claims remain the speaker's own self-reported statements โ worth staying a little skeptical about, rather than taking everything at face value just because it's labeled "free" and "most comprehensive."
That's really the same idea as "seeing what you're actually using and how much it costs" showing up from a different angle โ whether it's the tokens your own automated workflow burns through, or deciding whether to trust the numbers in a video, the starting point for judgment is the same: look clearly at what's actually happening, before deciding whether to invest anything.
Back to that first phone call. Once you actually get your first "call transcript auto-becomes-a-quote" skill running, you'll find that what really changes isn't how fast the quote gets written โ it's that you start habitually asking: can this repetitive piece of work become a skill I can call up anytime? And the next time your automated system is running several cycles a day, hundreds of times a week, can you actually answer how many tokens this "employee who never clocks out" is eating through every month?
You're probably not going to sit and watch this system run every single day going forward, but it's working โ and spending โ every single day regardless. Use AI Token King to lay out all that accumulated usage in plain view, so you're not the one blindsided when the bill lands at the end of the month. Try AI Token King free โ (https://aitokenking.com.tw/home)
FAQ
Are Codex and ChatGPT the same thing? Codex is OpenAI's product line built specifically for coding and automation tasks โ it can be installed as a standalone client, and you can also switch back to a regular ChatGPT conversation mode within the same interface. Both share the same account and subscription plan, but the interface and feature set are optimized for different kinds of tasks.
Can you actually build the automated workflow shown in the video with zero coding background? Nothing in the video's demo involves manually writing code. The approach is to describe what you need in plain language, review what Codex produces, and ask for changes when you're not satisfied. That lowers the barrier to entry, but you'll still need to spend time understanding what it actually did and reviewing whether what it generated is correct โ it isn't a zero learning-curve process.
How do you choose between "approve for me" and "full access" permission modes? If you're not confident the AI won't accidentally delete or alter an important file, start with "approve for me," so important actions get flagged for your confirmation. Once you're familiar with how it behaves and confident in your own backups, you can consider loosening that permission.
Is there really that big a difference between local automation and cloud scheduling? A local workflow is tied to your own computer โ it stops the moment you shut it down or lose connection. Once it's on a cloud schedule, the system keeps running on its set cycle even with your computer off. The difference comes down to whether you actually need "standing by 24/7" โ not every scenario requires going that far.
Is the claim about a 90-day money-back guarantee for landing your first client after joining the paid community credible? This is standard marketing language for course enrollment, and it's a one-sided promise from the speaker. This site hasn't verified the refund mechanism, the actual success rate, or the real-world outcomes โ readers should evaluate it on their own and not decide to join a paid community based solely on what's said in a video.
Source disclosure: This article is adapted from a teaching video published by the YouTube channel "ๆๅๅฉ | AI" (Harry Lee | AI), titled "ใๅ จ็ถฒๆๅ จใ2.5ๅฐๆๅฒ่ฉฉ็ด Codex ๅคงๅธซ่ชฒ๏ผๅพ้ถๆ้ 24ๅฐๆ็กไบบๅผๅฎ็ใไธไบบ AI ๅ ฌๅธใ" (roughly, "The Most Comprehensive 2.5-Hour Epic Codex Masterclass: Building a 24/7 Unattended One-Person AI Company From Scratch"), published 2026-08-31. Original link: https://www.youtube.com/watch?v=3fjO9siN-Nc. The community size and refund guarantee mentioned in this article are the speaker's own self-reported claims and have not been independently verified by this site.
Further reading: We've also covered what OpenAI Codex actually is and how it differs from ChatGPT and Claude Code for non-technical users, why AI tokens end up costing what they do once you break down how autoregressive generation burns through them in coding agents, five tested ways to cut token costs in AI coding workflows, and why software is increasingly moving to token-based pricing as AI agents become part of the procurement equation.
